| 佐々木 光 | D, 中間発表 | 知能システム制御 | 杉本 謙二☆, | 小笠原 司, | 松原 崇充 | |
|
title: Practical Bayesian Policy Search with Task-specific Latent Variable Priors
abstract: Policy search reinforcement learning algorithm is one of attractive method to learn a control policy for machine automation. While it has high performance in robot policy learning, it has not been applied to the real world since obtaining useful information for task execution in real world is limited. In this research, we propose a novel policy search algorithm to automate real-world task. This method introduces a latent variable suitable for a task by domain knowledge in policy search based on Bayesian inference. We proposed methods for objects manipulation and a waste crane in a garbage incineration plant, and confirmed the effectiveness of each task. language of the presentation: Japanese | ||||||
| 米野 尚斗 | M, 2回目発表 | 知能システム制御 | 杉本 謙二☆, | 小笠原 司, | 松原 崇充（特任准教授） | |
|
title:
Tactile perception of roughness due to tactile sensors and mechanical noise without slipping motion
abstract: Tactile sensing in robot control is an important element in the perception of the external world. However, the robot is always exposed to the risk of damage because mechanical contact with an object is essential to obtain tactile information. In this study, we propose a sensing system that enables the robot to recognize the texture of an object simply by touching it. The proposed system uses a PZT motor to apply vibrations to a soft structure of a biomimetic tactile sensor, and discriminates from changes in the vibration waveform. Experiments show that it is possible to classify the object's texture without controlling the robot's motion. language of the presentation: Japanese | ||||||
| OH HANBIT | M, 2回目発表 | 知能システム制御 | 杉本 謙二☆, | 小笠原 司, | 松原 崇充（特任准教授） | |
|
title: Multimodal Robust Imitation Learning with Noise injection
abstract: Imitation learning, whereby a robot learns policies to perform actions based on observing human demonstrations, is widely used in the field of robotics such as autonomous driving or industrial robot tasks. The State of the art method alleviates the covariate shift by adding a noise term to the demonstrator actions, inducing input error to robustify the learned policy. However, this approach assumes a task can be solved by a single optimal action called a unimodal policy. It cannot generate the desired action in a real-world scenario in which human demonstrations have multiple optimal actions. We propose an imitation learning method that uses an infinite overlapping mixture of Gaussian processes to obtain an injection noise that improves robustness and a multimodal policy that can learn multiple optimal actions at the same time. We validate the proposed method through simulations. language of the presentation: Japanese | ||||||
| 杉山 健太 | M, 2回目発表 | インタラクティブメディア設計学 | 加藤 博一, | 小笠原 司, | 神原 誠之, | 藤本 雄一郎 |
|
title: Fluid Simulation Using Particle Method Towards Reducing Drag on Swimmer
abstract: It is important to reduce drag on human from water in order to swim faster. A related work suggested the gravity position of swimmer should be higher in order to reduce drag on swimmer because the underwater frontal projected area goes smaller. However, they cut water surface as plane. In real-world environment, water covers the head. We propose the method to simulate fluid behavior using boundary particles even if the shape of 3D model mesh change like human swimming. We find out whether the drag go smaller or not if underwater frontal projected area reduces. In this colloquium B presentation, I present the fluid behavior with human running motion and the comparison between open-source simulator and our simulator in cube translation scene. language of the presentation: English | ||||||